michaelfeil/infinity v3.0.0-alpha.6

Infinity v3.0.0-alpha.6

on GitHub

Changes since alpha.5:

• Add packed tensor embedding output with shape and dtype metadata: MessagePack binary or JSON base64, including late-interaction token matrices (#725).
• Accelerate LibTorch BF16 Qwen3 and Llama prefill (#722).
• Fuse LibTorch dynamic FP8 prefill and default its dense decoder scope to all-linear; dynamic FP8 remains opt-in (#721, #724).
• Fix loading of normalized Ministral3 checkpoints with independent attention dimensions (#723).
• Adopt Infinity Restricted MIT License 1.0 for covered V3 contributions, preserving earlier grants and third-party terms (#720).

Images publish asynchronously after the release builds and startup checks pass:

docker pull ghcr.io/michaelfeil/infinity_v3:nvidia-3.0.0-alpha.6
docker pull ghcr.io/michaelfeil/infinity_v3:cpu-3.0.0-alpha.6

Image builds · Tensor output guide · Licensing · Full changelog